vllm-cpu 0.9.2.post2__cp311-cp311-manylinux_2_17_aarch64.whl

vllm/compilation/cuda_piecewise_backend.py

@@ -0,0 +1,218 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import dataclasses
+from contextlib import ExitStack
+from typing import Any, Callable, Optional
+from unittest.mock import patch
+
+import torch
+import torch.fx as fx
+
+import vllm.envs as envs
+from vllm.compilation.backends import VllmBackend
+from vllm.compilation.counter import compilation_counter
+from vllm.compilation.monitor import end_monitoring_torch_compile
+from vllm.config import VllmConfig
+from vllm.forward_context import get_forward_context
+from vllm.logger import init_logger
+from vllm.utils import weak_ref_tensors
+
+logger = init_logger(__name__)
+
+
+@dataclasses.dataclass
+class ConcreteSizeEntry:
+    runtime_shape: int
+    need_to_compile: bool  # the size is in compile_sizes
+    use_cudagraph: bool  # the size is in cudagraph_capture_sizes
+
+    compiled: bool = False
+    runnable: Callable = None  # type: ignore
+    num_finished_warmup: int = 0
+    cudagraph: Optional[torch.cuda.CUDAGraph] = None
+    output: Optional[Any] = None
+
+    # for cudagraph debugging, track the input addresses
+    # during capture, and check if they are the same during replay
+    input_addresses: Optional[list[int]] = None
+
+
+class CUDAPiecewiseBackend:
+
+    def __init__(self, graph: fx.GraphModule, vllm_config: VllmConfig,
+                 graph_pool: Any, piecewise_compile_index: int,
+                 total_piecewise_compiles: int, sym_shape_indices: list[int],
+                 compiled_graph_for_general_shape: Callable,
+                 vllm_backend: VllmBackend):
+        """
+        The backend for piecewise compilation.
+        It mainly handles the compilation and cudagraph capturing.
+
+        We will compile `self.graph` once for the general shape,
+        and then compile for different shapes specified in
+        `compilation_config.compile_sizes`.
+
+        Independently, we will capture cudagraph for different shapes.
+
+        If a shape needs both compilation and cudagraph, we will
+        compile it first, and then capture cudagraph.
+        """
+        self.graph = graph
+        self.vllm_config = vllm_config
+        self.compilation_config = vllm_config.compilation_config
+        self.graph_pool = graph_pool
+        self.piecewise_compile_index = piecewise_compile_index
+        self.total_piecewise_compiles = total_piecewise_compiles
+        self.vllm_backend = vllm_backend
+
+        self.is_first_graph = piecewise_compile_index == 0
+        self.is_last_graph = (
+            piecewise_compile_index == total_piecewise_compiles - 1)
+
+        self.compile_sizes: set[int] = set(
+            self.compilation_config.compile_sizes)
+        self.cudagraph_capture_sizes: set[int] = set(
+            self.compilation_config.cudagraph_capture_sizes
+        ) if self.compilation_config.use_cudagraph else set()
+
+        self.first_run_finished = False
+
+        self.compiled_graph_for_general_shape = compiled_graph_for_general_shape  # noqa
+
+        self.sym_shape_indices = sym_shape_indices
+
+        self.is_debugging_mode = envs.VLLM_LOGGING_LEVEL == "DEBUG"
+
+        # the entries for different shapes that we need to either
+        # compile or capture cudagraph
+        self.concrete_size_entries: dict[int, ConcreteSizeEntry] = {}
+
+        # to_be_compiled_sizes tracks the remaining sizes to compile,
+        # and updates during the compilation process, so we need to copy it
+        self.to_be_compiled_sizes: set[int] = self.compile_sizes.copy()
+        for shape in self.compile_sizes.union(self.cudagraph_capture_sizes):
+            self.concrete_size_entries[shape] = ConcreteSizeEntry(
+                runtime_shape=shape,
+                need_to_compile=shape in self.compile_sizes,
+                use_cudagraph=shape in self.cudagraph_capture_sizes,
+            )
+
+    def check_for_ending_compilation(self):
+        if self.is_last_graph and not self.to_be_compiled_sizes:
+            # no specific sizes to compile
+            # save the hash of the inductor graph for the next run
+            self.vllm_backend.compiler_manager.save_to_file()
+            end_monitoring_torch_compile(self.vllm_config)
+
+    def __call__(self, *args) -> Any:
+        if not self.first_run_finished:
+            self.first_run_finished = True
+            self.check_for_ending_compilation()
+            return self.compiled_graph_for_general_shape(*args)
+
+        runtime_shape = args[self.sym_shape_indices[0]]
+        if runtime_shape not in self.concrete_size_entries:
+            # we don't need to do anything for this shape
+            return self.compiled_graph_for_general_shape(*args)
+
+        entry = self.concrete_size_entries[runtime_shape]
+
+        if entry.runnable is None:
+            entry.runnable = self.compiled_graph_for_general_shape
+
+        if entry.need_to_compile and not entry.compiled:
+            entry.compiled = True
+            self.to_be_compiled_sizes.remove(runtime_shape)
+            # args are real arguments
+            entry.runnable = self.vllm_backend.compiler_manager.compile(
+                self.graph,
+                args,
+                self.compilation_config.inductor_compile_config,
+                self.compilation_config,
+                graph_index=self.piecewise_compile_index,
+                num_graphs=self.total_piecewise_compiles,
+                runtime_shape=runtime_shape)
+
+            # finished compilations for all required shapes
+            if self.is_last_graph and not self.to_be_compiled_sizes:
+                self.check_for_ending_compilation()
+
+        # Skip CUDA graphs if this entry doesn't use them OR
+        # if we're supposed to skip them globally
+        skip_cuda_graphs = get_forward_context().skip_cuda_graphs
+        if not entry.use_cudagraph or skip_cuda_graphs:
+            return entry.runnable(*args)
+
+        if entry.cudagraph is None:
+            if entry.num_finished_warmup < self.compilation_config.cudagraph_num_of_warmups:  # noqa
+                entry.num_finished_warmup += 1
+                if self.is_first_graph:
+                    logger.debug(
+                        "Warming up %s/%s for shape %s",
+                        entry.num_finished_warmup,
+                        self.compilation_config.cudagraph_num_of_warmups,
+                        runtime_shape)
+                return entry.runnable(*args)
+
+            if self.is_first_graph:
+                # Since we capture cudagraph for many different shapes and
+                # capturing is fast, we don't need to log it for every shape.
+                # We only log it in the debug mode.
+                logger.debug("Capturing a cudagraph for shape %s",
+                             runtime_shape)
+
+            input_addresses = [
+                x.data_ptr() for x in args if isinstance(x, torch.Tensor)
+            ]
+            entry.input_addresses = input_addresses
+            cudagraph = torch.cuda.CUDAGraph()
+
+            with ExitStack() as stack:
+                if not self.is_first_graph:
+                    # during every model forward, we will capture
+                    # many pieces of cudagraphs (roughly one per layer).
+                    # running gc again and again across layers will
+                    # make the cudagraph capture very slow.
+                    # therefore, we only run gc for the first graph,
+                    # and disable gc for the rest of the graphs.
+                    stack.enter_context(patch("gc.collect", lambda: None))
+                    stack.enter_context(
+                        patch("torch.cuda.empty_cache", lambda: None))
+
+                # mind-exploding: carefully manage the reference and memory.
+                with torch.cuda.graph(cudagraph, pool=self.graph_pool):
+                    # `output` is managed by pytorch's cudagraph pool
+                    output = entry.runnable(*args)
+                    if self.is_last_graph:
+                        # by converting it to weak ref,
+                        # the original `output` will immediately be released
+                        # to save memory. It is only safe to do this for
+                        # the last graph, because the output of the last graph
+                        # will not be used by any other cuda graph.
+                        output = weak_ref_tensors(output)
+
+            # here we always use weak ref for the output
+            # to save memory
+            entry.output = weak_ref_tensors(output)
+            entry.cudagraph = cudagraph
+
+            compilation_counter.num_cudagraph_captured += 1
+
+            # important: we need to return the output, rather than
+            # the weak ref of the output, so that pytorch can correctly
+            # manage the memory during cuda graph capture
+            return output
+
+        if self.is_debugging_mode:
+            # check if the input addresses are the same
+            new_input_addresses = [
+                x.data_ptr() for x in args if isinstance(x, torch.Tensor)
+            ]
+            assert new_input_addresses == entry.input_addresses, (
+                "Input addresses for cudagraphs are different during replay."
+                f" Expected {entry.input_addresses}, got {new_input_addresses}"
+            )
+
+        entry.cudagraph.replay()
+        return entry.output

vllm/compilation/decorators.py

@@ -0,0 +1,250 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import inspect
+from typing import Callable, Optional, TypeVar, Union, overload
+from unittest.mock import patch
+
+import torch
+import torch.nn as nn
+from torch._dynamo.symbolic_convert import InliningInstructionTranslator
+
+from vllm.compilation.counter import compilation_counter
+from vllm.compilation.wrapper import TorchCompileWrapperWithCustomDispatcher
+from vllm.config import CompilationLevel, VllmConfig
+from vllm.logger import init_logger
+from vllm.sequence import IntermediateTensors
+from vllm.utils import supports_dynamo
+
+from .monitor import start_monitoring_torch_compile
+
+logger = init_logger(__name__)
+
+_T = TypeVar("_T", bound=type[nn.Module])
+
+
+@overload
+def support_torch_compile(
+    *,
+    dynamic_arg_dims: Optional[dict[str, Union[int, list[int]]]],
+) -> Callable[[_T], _T]:
+    ...
+
+
+@overload
+def support_torch_compile(cls: _T) -> _T:
+    ...
+
+
+def support_torch_compile(
+    cls: Optional[_T] = None,
+    *,
+    dynamic_arg_dims: Optional[dict[str, Union[int, list[int]]]] = None,
+) -> Union[Callable[[_T], _T], _T]:
+    """
+    A decorator to add support for compiling the forward method of a class.
+
+    Usage 1: use directly as a decorator without arguments:
+
+    ```python
+    @support_torch_compile
+    class MyModel(nn.Module):
+        def forward(self, x: torch.Tensor, y: Optional[torch.Tensor]):
+            ...
+    ```
+
+    Usage 2: use as a decorator with arguments:
+
+    ```python
+    @support_torch_compile(dynamic_arg_dims={"x": 0, "y": 0})
+    class MyModel(nn.Module):
+        def forward(self, x: torch.Tensor, y: Optional[torch.Tensor]):
+            ...
+    ```
+
+    `dynamic_arg_dims` is a dictionary that maps argument names to the dynamic
+    dimensions of the argument. The dynamic dimensions can be either a single
+    integer or a list of integers.
+
+    if `dynamic_arg_dims` is `None`, it is inferred from the type annotation
+    of the `forward` method, based on the following default rules:
+
+    - if the argument is annotated as `torch.Tensor` or
+        `Optional[torch.Tensor]`, the first dimension will be
+        marked as dynamic.
+    - if the argument is annotated as `IntermediateTensors`, the first
+        dimension of all the tensors in the intermediate tensors
+        will be marked as dynamic.
+
+    During runtime, when we actually mark dimensions of tensors,
+     it depends on the value of arguments:
+
+    - if it is a single integer (can be negative), the corresponding dimension 
+        of the argument will be marked as dynamic.
+    - if it is `None`, ignored.
+    - if it is `IntermediateTensors`, all the tensors in the intermediate
+        tensors will be marked as dynamic.
+    - otherwise, it will raise an error.
+
+    NOTE: if an argument is `None`, it should always be passed as `None` during
+    the lifetime of the model, otherwise, it cannot be captured as a single
+    computation graph.
+    """
+
+    def cls_decorator_helper(cls: _T) -> _T:
+        # helper to pass `dynamic_arg_dims`` to `_support_torch_compile``
+        # to avoid too much indentation for `_support_torch_compile``
+        if not hasattr(cls, 'forward'):
+            raise TypeError("decorated class should have a forward method.")
+        sig = inspect.signature(cls.forward)
+        inferred_dynamic_arg_dims = dynamic_arg_dims
+        if inferred_dynamic_arg_dims is None:
+            inferred_dynamic_arg_dims = {}
+            for k, v in sig.parameters.items():
+                if v.annotation in [
+                        torch.Tensor, Optional[torch.Tensor],
+                        IntermediateTensors, Optional[IntermediateTensors]
+                ]:
+                    inferred_dynamic_arg_dims[k] = 0
+
+            logger.debug(("Inferred dynamic dimensions for "
+                          "forward method of %s: %s"), cls,
+                         list(inferred_dynamic_arg_dims.keys()))
+
+        if len(inferred_dynamic_arg_dims) == 0:
+            raise ValueError(
+                "No dynamic dimensions found in the forward method of "
+                f"{cls}. Please provide dynamic_arg_dims explicitly.")
+
+        for k in inferred_dynamic_arg_dims:
+            if k not in sig.parameters:
+                raise ValueError(
+                    f"Argument {k} not found in the forward method of {cls}")
+        return _support_torch_compile(cls, inferred_dynamic_arg_dims)
+
+    if cls is not None:
+        # use `support_torch_compile` as a decorator without arguments
+        assert isinstance(cls, type)
+        return cls_decorator_helper(cls)
+
+    return cls_decorator_helper
+
+
+def _support_torch_compile(
+    cls: _T,
+    dynamic_arg_dims: dict[str, Union[int, list[int]]],
+) -> _T:
+    """
+    A decorator to add support for compiling the forward method of a class.
+    """
+    if TorchCompileWrapperWithCustomDispatcher in cls.__bases__:
+        # support decorating multiple times
+        return cls
+
+    # take care of method resolution order
+    # make sure super().__init__ is called on the base class
+    #  other than TorchCompileWrapperWithCustomDispatcher
+    cls.__bases__ = cls.__bases__ + (TorchCompileWrapperWithCustomDispatcher, )
+
+    old_init = cls.__init__
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = '', **kwargs):
+        old_init(self, vllm_config=vllm_config, prefix=prefix, **kwargs)
+        self.vllm_config = vllm_config
+        # for CompilationLevel.DYNAMO_AS_IS , the upper level model runner
+        # will handle the compilation, so we don't need to do anything here.
+        self.do_not_compile = \
+            vllm_config.compilation_config.level in [
+            CompilationLevel.NO_COMPILATION, CompilationLevel.DYNAMO_AS_IS
+        ] or not supports_dynamo()
+        if self.do_not_compile:
+            return
+        compilation_counter.num_models_seen += 1
+        TorchCompileWrapperWithCustomDispatcher.__init__(
+            self, compilation_level=vllm_config.compilation_config.level)
+
+    cls.__init__ = __init__
+
+    def __call__(self, *args, **kwargs):
+        # torch.compiler.is_compiling() means we are inside the compilation
+        # e.g. TPU has the compilation logic in model runner, so we don't
+        # need to compile the model inside.
+        if self.do_not_compile or torch.compiler.is_compiling():
+            return self.forward(*args, **kwargs)
+
+        # the first compilation needs to have dynamic shapes marked
+        if len(self.compiled_codes) < 1:
+            sig = inspect.signature(self.__class__.forward)
+            bound_args = sig.bind(self, *args, **kwargs)
+            bound_args.apply_defaults()
+            for k, dims in dynamic_arg_dims.items():
+                arg = bound_args.arguments.get(k)
+                if arg is not None:
+                    dims = [dims] if isinstance(dims, int) else dims
+                    if isinstance(arg, torch.Tensor):
+                        # In case dims is specified with negative indexing
+                        dims = [
+                            arg.ndim + dim if dim < 0 else dim for dim in dims
+                        ]
+                        torch._dynamo.mark_dynamic(arg, dims)
+                    elif isinstance(arg, IntermediateTensors):
+                        for tensor in arg.tensors.values():
+                            # In case dims is specified with negative indexing
+                            dims = [
+                                tensor.ndim + dim if dim < 0 else dim
+                                for dim in dims
+                            ]
+                            torch._dynamo.mark_dynamic(tensor, dims)
+                    else:
+                        raise ValueError(
+                            "Unsupported dynamic dimensions"
+                            f" {dims} for argument {k} with type {type(arg)}.")
+            # here, it is the starting point of the `torch.compile` process
+            start_monitoring_torch_compile(self.vllm_config)
+            logger.debug("Start compiling function %s",
+                         self.original_code_object)
+
+        # if we don't use custom dispatcher, we can directly call the
+        # compiled function and let torch.compile handle the dispatching,
+        # with the overhead of guard evaluation and recompilation.
+        if len(self.compiled_codes) < 1 or not self.use_custom_dispatcher:
+            # it seems Dynamo reuse the compilation across instances,
+            # while we need to make sure the compiled code is not reused.
+            # we need to control all the compilation of the model.
+            torch._dynamo.eval_frame.remove_from_cache(
+                self.original_code_object)
+
+            # collect all relevant files traced by Dynamo,
+            # so that the compilation cache can trigger re-compilation
+            # properly when any of these files change.
+
+            # 1. the file containing the top-level forward function
+            self.vllm_config.compilation_config.traced_files.add(
+                self.original_code_object.co_filename)
+
+            # 2. every time Dynamo sees a function call, it will inline
+            # the function by calling InliningInstructionTranslator.inline_call
+            # we hijack this function to know all the functions called
+            # during Dynamo tracing, and their corresponding files
+            inline_call = InliningInstructionTranslator.inline_call
+
+            def patched_inline_call(parent, func, args, kwargs):
+                code = func.get_code()
+                self.vllm_config.compilation_config.traced_files.add(
+                    code.co_filename)
+                return inline_call(parent, func, args, kwargs)
+
+            with patch.object(InliningInstructionTranslator, 'inline_call',
+                              patched_inline_call):
+                output = self.compiled_callable(*args, **kwargs)
+            return output
+
+        # usually, capturing the model once is enough, and then we can
+        # dispatch to the compiled code directly, without going through
+        # the Dynamo guard mechanism.
+        with self.dispatch_to_code(0):
+            model_output = self.forward(*args, **kwargs)
+            return model_output
+
+    cls.__call__ = __call__
+    return cls

vllm/compilation/fix_functionalization.py

@@ -0,0 +1,191 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import operator
+from collections.abc import Iterable
+from typing import Optional, Union
+
+import torch
+from torch._higher_order_ops.auto_functionalize import auto_functionalized
+
+from vllm.logger import init_logger
+
+from .fx_utils import is_func
+from .vllm_inductor_pass import VllmInductorPass
+
+logger = init_logger(__name__)
+
+
+class FixFunctionalizationPass(VllmInductorPass):
+    """
+    This pass defunctionalizes certain nodes to avoid redundant tensor copies.
+    After this pass, DCE (dead-code elimination) should never be run,
+    as de-functionalized nodes may appear as dead code.
+
+    To add new nodes to defunctionalize, add to the if-elif chain in __call__.
+    """
+
+    def __call__(self, graph: torch.fx.Graph):
+        self.begin()
+        self.dump_graph(graph, "before_fix_functionalization")
+
+        self.nodes_to_remove: list[torch.fx.Node] = []
+        count = 0
+        for node in graph.nodes:
+            if not is_func(node, auto_functionalized):
+                continue  # Avoid deep if-elif nesting
+
+            kwargs = node.kwargs
+            at_target = node.args[0]
+
+            if at_target == torch.ops._C.rotary_embedding.default:
+                query = kwargs['query']
+                mm_node = query.args[0].args[0]
+
+                # rotary_embedding is a special case: the two mutating inputs
+                # are query and key, which are slices of mm_node.
+                # While functionalized, results at[1] and at[2] are scattered
+                # back into mm_node. After de-functionalization, we can just
+                # use mm_node directly.
+                for idx, user in self.getitem_users(node).items():
+                    for user_of_getitem in user.users:
+                        if is_func(user_of_getitem,
+                                   torch.ops.aten.slice_scatter.default):
+                            user_of_getitem.replace_all_uses_with(mm_node)
+                            self._remove(user_of_getitem)
+                    self._remove(user)
+
+                self.insert_defunctionalized(graph, node)
+                self._remove(node)
+
+            # rms_norm replacements avoid the most copies for LLaMa.
+            elif at_target == torch.ops._C.fused_add_rms_norm.default:
+                mutated_args = {1: 'input', 2: 'residual'}
+                self.defunctionalize(graph, node, mutated_args)
+            elif at_target == torch.ops._C.fused_add_rms_norm_static_fp8_quant.default:  # noqa: E501
+                mutated_args = {1: 'result', 2: 'residual'}
+                self.defunctionalize(graph, node, mutated_args)
+            elif at_target == torch.ops._C.rms_norm_dynamic_per_token_quant.default:  # noqa: E501
+                mutated_args = {1: 'result', 2: 'scale', 3: 'residual'}
+                self.defunctionalize(graph, node, mutated_args)
+            elif at_target in [
+                    torch.ops._C.rms_norm.default,
+                    torch.ops._C.rms_norm_static_fp8_quant.default,
+            ]:
+                mutated_args = {1: 'result'}
+                self.defunctionalize(graph, node, mutated_args)
+            # For some reason we need to specify the args for both
+            # silu_and_mul and silu_and_mul_quant. The kwargs
+            # pathway gets the wrong answer.
+            elif at_target == torch.ops._C.silu_and_mul.default:
+                mutated_args = {1: 'result'}
+                self.defunctionalize(graph,
+                                     node,
+                                     mutated_args,
+                                     args=('result', 'input'))
+            elif at_target == torch.ops._C.silu_and_mul_quant.default:
+                mutated_args = {1: 'result'}
+                self.defunctionalize(graph,
+                                     node,
+                                     mutated_args,
+                                     args=('result', 'input', 'scale'))
+            else:
+                continue  # skip the count
+
+            count += 1
+
+        self.dump_graph(graph, "before_fix_functionalization_cleanup")
+
+        # Remove the nodes all at once
+        count_removed = len(self.nodes_to_remove)
+        for node in self.nodes_to_remove:
+            graph.erase_node(node)
+
+        logger.debug("De-functionalized %s nodes, removed %s nodes", count,
+                     count_removed)
+        self.dump_graph(graph, "after_fix_functionalization")
+        self.end_and_log()
+
+    def _remove(self, node_or_nodes: Union[torch.fx.Node,
+                                           Iterable[torch.fx.Node]]):
+        """
+        Stage a node (or nodes) for removal at the end of the pass.
+        """
+        if isinstance(node_or_nodes, torch.fx.Node):
+            self.nodes_to_remove.append(node_or_nodes)
+        else:
+            self.nodes_to_remove.extend(node_or_nodes)
+
+    def defunctionalize(self,
+                        graph: torch.fx.Graph,
+                        node: torch.fx.Node,
+                        mutated_args: dict[int, Union[torch.fx.Node, str]],
+                        args: Optional[tuple[Union[torch.fx.Node, str],
+                                             ...]] = None):
+        """
+        De-functionalize a node by replacing it with a call to the original.
+        It also replaces the getitem users with the mutated arguments.
+        See replace_users_with_mutated_args and insert_defunctionalized.
+        """
+        self.replace_users_with_mutated_args(node, mutated_args)
+        self.insert_defunctionalized(graph, node, args=args)
+        self._remove(node)
+
+    def replace_users_with_mutated_args(self, node: torch.fx.Node,
+                                        mutated_args: dict[int,
+                                                           Union[torch.fx.Node,
+                                                                 str]]):
+        """
+        Replace all getitem users of the auto-functionalized node with the
+        mutated arguments.
+        :param node: The auto-functionalized node
+        :param mutated_args: The mutated arguments, indexed by getitem index.
+        If the value of an arg is a string, `node.kwargs[arg]` is used.
+        """
+        for idx, user in self.getitem_users(node).items():
+            arg = mutated_args[idx]
+            arg = node.kwargs[arg] if isinstance(arg, str) else arg
+            user.replace_all_uses_with(arg)
+            self._remove(user)
+
+    def getitem_users(self, node: torch.fx.Node) -> dict[int, torch.fx.Node]:
+        """
+        Returns the operator.getitem users of the auto-functionalized node,
+        indexed by the index they are getting.
+        """
+        users = {}
+        for user in node.users:
+            if is_func(user, operator.getitem):
+                idx = user.args[1]
+                users[idx] = user
+        return users
+
+    def insert_defunctionalized(self,
+                                graph: torch.fx.Graph,
+                                node: torch.fx.Node,
+                                args: Optional[tuple[Union[torch.fx.Node, str],
+                                                     ...]] = None):
+        """
+        Insert a new defunctionalized node into the graph before node.
+        If one of the kwargs is 'out', provide args directly,
+        as node.kwargs cannot be used.
+        See https://github.com/pytorch/pytorch/blob/a00faf440888ffb724bad413f329a49e2b6388e7/torch/_inductor/lowering.py#L351
+
+        :param graph: Graph to insert the defunctionalized node into
+        :param node: The auto-functionalized node to defunctionalize
+        :param args: If we cannot use kwargs, specify args directly.
+        If an arg is a string, `node.kwargs[arg]` is used.
+        """  # noqa: E501
+        assert is_func(node, auto_functionalized), \
+            f"node must be auto-functionalized, is {node} instead"
+
+        # Create a new call to the original function
+        with graph.inserting_before(node):
+            function = node.args[0]
+            if args is None:
+                graph.call_function(function, kwargs=node.kwargs)
+            else:
+                # Args passed as strings refer to items in node.kwargs
+                args = tuple(node.kwargs[arg] if isinstance(arg, str) else arg
+                             for arg in args)
+                graph.call_function(function, args=args)